Meta Platforms Inc. (META.US) is actively working to expand the application boundaries of its custom chips, according to Chief Financial Officer Susan Li, despite the company's recent significant deals with leading chip manufacturers. She explained that due to the highly specialized nature of some of Meta's workloads, in-house developed chips can better accommodate specific internal algorithm requirements. Currently, Meta has achieved large-scale deployment of custom chips within its core ranking and recommendation systems. The future strategic focus will be to gradually extend this capability into the domain of training artificial intelligence models. Although not a traditional cloud computing provider, Meta operates one of the world's largest data center networks dedicated to training and running AI models. In recent weeks, the company has entered into multiple major agreements with industry leader NVIDIA (NVDA.US) and competitor AMD (AMD.US) to procure chips and equipment for supporting AI workloads. Concurrently, the social media giant continues to advance the development of its internal AI processors. Susan Li emphasized that Meta is sourcing different types of chips to suit diverse task requirements. "Based on current understanding and practical needs, we are systematically evaluating the most suitable chip solution for each application scenario," she stated, noting that "custom chips remain a central pillar of this strategic framework." This announcement signifies a critical advancement phase for Meta's in-house chip project (MTIA). Since first revealing the MTIA plan in 2023, Meta's initial development focus was primarily on the inference phase, aiming to enhance the computational efficiency of Facebook and Instagram recommendation systems and reduce reliance on NVIDIA's general-purpose GPUs. With the explosion of generative AI, Meta's demand for computing power has grown exponentially, making a focus solely on inference insufficient to support its large language model strategy. Susan Li's latest statement sends a clear signal to the market: despite industry skepticism regarding the high barriers to developing top-tier AI training chips, Meta remains steadfast in viewing "self-developed training chips" as the ultimate goal of its infrastructure transformation. However, the path to computational self-reliance is not without challenges. Recent market reports suggest that Meta has encountered certain technical obstacles in developing its most advanced training chips, with rumors even indicating potential adjustments to the timeline for some high-performance projects. To balance immediate high-performance computing gaps with long-term in-house development goals, Meta has adopted a flexible, diversified supply strategy. On one hand, reports indicate Meta has reached an agreement with Google to rent its TPU resources to accelerate the development of large models in the current phase. On the other hand, the company maintains deep procurement relationships with NVIDIA. Susan Li's emphasis on "gradually expanding over time" suggests Meta will pursue a steady, step-by-step transition model—achieving breakthroughs in specific, customized tasks first, before ultimately conquering the computing challenges of general large model training. From an industry perspective, Meta's chip development journey reflects a common logic among hyperscale cloud players in the AI era: full-stack in-house development. By deeply integrating chip architecture with proprietary models like Llama, Meta not only aims to significantly reduce long-term hardware procurement and energy costs but also to avoid being constrained by supply chain fluctuations. Although the leap from recommendation system inference to complex model training presents immense architectural challenges, Meta, leveraging its vast application scenarios and substantial cash flow, is attempting to redefine the balance of power between internet giants and hardware suppliers.
Comments